I am curious about RAC with the stock v8 apps. I gather from reading threads here that it is down somewhat from v7; has anyone some kind of numbers for this (percentage-wise, I mean). I do know the v8 apps do more work than the v7 did, and run longer, in consequence. But I don't know how that translates to RAC.

Also, anyone have any numbers about how much (if anything) the new Lunatics 0.44 apps improve that stock v8 performance?

When I was running Lunatics late last year, my two machines were producing around 125/130K per day of RAC; I have been running v8 only stock (until just an hour ago, when I finally installed 0.44 on my machines), and my RAC is right around 100K. So what can I expect from the upgrade?

(I run both MB (CPU and GPU), and AP (GPU only), and I just take what I get from the servers. I am also running the same number of WUs/GPU now - 3 - as I was before v8).).

Thanks for any info!

---

I do know the v8 apps do more work than the v7 did, and run longer, in consequence. But I don't know how that translates to RAC.

RAC will be significantly lower, but it's still hard to say by how much. I was running optimised CPU & GPU v7 applications, and my RAC was around 49,000.
With v8, I'm running Lunatics again, and I'm using the AVX application on my i7, and the SSE3 application on my Core 2 Duo.
On the C2D it is significantly slower, the AVX on the i7 is doing the work in roughly the same time as the v7 AVX application.
However there isn't an optimised CUDA application yet, so that alone will be the cause of a significant drop in RAC.

My RAC used to be around 49,000. Since the drop in RAC from the change over period and finally getting the GPU application my RAC has been climbing, but it has slowed greatly over the last few weeks & looks like it might finally have peaked at around 43,000.

Also, anyone have any numbers about how much (if anything) the new Lunatics 0.44 apps improve that stock v8 performance?

For non AVX work my C2D take almost an hour longer for the longer running WUs than v7 did with the SSSE3 application.
With the current v8 AVX application on my Win7 i7 the run times are pretty much on par with the v7 AVX application.

---

Grant
Darwin NT

However there isn't an optimised CUDA application yet, so that alone will be the cause of a significant drop in RAC.

For both MB v7 and MB v8, there is no difference in 'optimisation' between the stock CUDA app and the Lunatics CUDA app - they are the same apps (Berkeley doesn't have the manpower to write their own CUDA apps, so without 'borrowing' the Lunatics app, there would be no stock CUDA app). The only difference is that you can choose which CUDA version to run from the get-go with Lunatics.

I'm monitoring the RAC on a single GTX 750 Ti running MB continuously on GPU only (it has a high factory overclock, but hasn't been tweaked beyond that). In the run up to Christmas, RAC was fluctuating between 12,500 and 13,000 - the highest value reached was 13,027.

That was in a Windows XP host supported by a Q6600. That machine died in the winter storms (power supply problems), so the same card is now running under Windows 7 supported by an i5 - and in a PCIe x4 slot, because there's a new card in the primary slot.

v8 RAC is only now just about stabilising, at round about 11,500 - highest value seen so far (two days ago) is 11,637.

So that's about a 10% reduction in RAC. How much of that is due to the v7/v8 transition, and how much to the different hardware and OS environment, I'll leave you to discuss.

However there isn't an optimised CUDA application yet, so that alone will be the cause of a significant drop in RAC.

For both MB v7 and MB v8, there is no difference in 'optimisation' between the stock CUDA app and the Lunatics CUDA app - they are the same apps (Berkeley doesn't have the manpower to write their own CUDA apps, so without 'borrowing' the Lunatics app, there would be no stock CUDA app). The only difference is that you can choose which CUDA version to run from the get-go with Lunatics.

Ok, that would explain why the drop in RAC isn't nearly as much as i thought it would be.
It looks like the RAC hit will probably be between 13-18% or so. A bit of that will be the loss of the SSSE3 optimisations that the v7 client had, although the output of the C2D CPU is only a small portion of total output. So maybe a 11-16% drop. Will take a few more weeks to see just where the RAC ends up.

---

Grant
Darwin NT

It looks like the RAC hit will probably be between 13-18% or so. A bit of that will be the loss of the SSSE3 optimisations that the v7 client had, although the output of the C2D CPU is only a small portion of total output. So maybe a 11-16% drop. Will take a few more weeks to see just where the RAC ends up.

A week later & my RAC is still (although even more slowly than it was) rising.
The final drop in RAC may only be in the 8-12% range. Not good, but no where nearly as bad as I was expecting it to be.

---

Grant
Darwin NT

My RAC was about 28,000 around the first of the year and it went as low as 14,000 at the end of January while waiting for a GPU/Linux app. Six weeks later I am now at 22,000 and I think if I am lucky I might hit 24,000 if the APs pick up a bit Overall I think will be down 12-15% as result of the V7 to V8 transition.

It looks like the RAC hit will probably be between 13-18% or so. A bit of that will be the loss of the SSSE3 optimisations that the v7 client had, although the output of the C2D CPU is only a small portion of total output. So maybe a 11-16% drop. Will take a few more weeks to see just where the RAC ends up.

A week later & my RAC is still (although even more slowly than it was) rising.
The final drop in RAC may only be in the 8-12% range. Not good, but no where nearly as bad as I was expecting it to be.

8-12% would seem to reflect the precision changes Eric made (mainly Gaussian related), which mandate that certain codepaths with certain compilers had to be switched from fast math to precise mode. At present in my own Cuda builds that use ms Visual STudio Compilers, it's the case of shotgun setting everything to fp:precise for the CPU side. That's something where as the accuracy limitations of each section of code become clear, some areas (probably the majority) can be put back, while remaining precision sensitive areas become candidates for optimisation.

As to why the increased, mainly Gaussian related, precision demands in v8, well reviewing the changes indicates preparation for processing data from other telescopes. I suspect some reasonable portion of these would include reobservation data of specific points in the sky, therefore solving VLAR performance problems, and careful treatment of those precios WoW-like Gaussians will become a high priority soon, perhaps pushing our cheap and cheerful shorty VHAR focus aside.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Having been running the optimized apps (AVX, SSE4, and AMD HD5) on my top two crunchers for a while now, my RAC actually seems higher than it was with the optimized apps under version 7. Most notably on my 4790/R9280x box, I don't think RAC ever hit 30k before and now it hasn't dipped below that for at least a couple weeks. Granted, I do understand that AP7 units give a lot more credit than MB8. Has there been more AP7 work than usual going out lately?